The field of research

Process: both. The talk is the working process; whatever is born goes through create-element or the hand's own agent (build). Core step 701; the owner, 2026-11-08: «любое исследование должно быть подкреплено реальными тестами, и осмысленными анализами … сложные продукты могут проходить стадию исследования в течение продолжительного времени и поэтому сохранять контекст для таких задач становится важной задачей … И тем более если понадобится перекрёстный анализ». Step 673: this skill gives the facts of the field; which shape the work takes, the method of shape-the-need decides on the node's facts.

Why it matters

A research told from documentation is a retelling. A research proven on this computer — the SDK installed, a server and a client started, the API called with a test key — is evidence, and evidence that is kept is what the next research and the next comparison stand on.

What the words can mean

• The words «this project» have three readings: an outside product with a link or a name — a research; this node or Fractera — npm run contacts and npm run card, answered from them (CLAUDE.md §4, «Know yourself»); an element of the node — its card and a question to its agent (ask-element). No link and no name — that is the one question.
• The three usual questions — how it works, what it solves, how to integrate — need three kinds of evidence: the mechanism (documentation, the repository, the product itself in a test account the person confirmed); the value (its own pages, prices, dated reviews); the integration (API, SDK, MCP, webhooks, sign-in, limits, terms of use) and a live probe with a test key.
• The goal behind the question — to integrate, to build an analogue, to buy, to write about it, to decide — sets the depth.

The facts the forces act on

• Primary sources only. A claim about a foreign service comes from its own pages, its repository or its terms, with the date it was read (CLAUDE.md §6). A site that forbids study gives only the person's screenshots, its inner pages only a test account the person confirmed (npm run site-rules).
• A live probe needs this computer: the SDK installed, a server and a client started, a call with a test key that lives in the environment of the hand that probes, never in the node's .env.local. Its scripts stay in that hand's git and run again later.
• What a chat already does — Anthropic's own pages, read 2026-11-08: Claude.ai's Research runs agentic searches that build on each other, with citations (anthropic.com/news/research, 2025-04-15); Integrations connect tools through remote MCP servers, and a research runs up to 45 minutes (anthropic.com/news/integrations, 2025-05-01). A one-off question with nothing to keep is answered well there, or here in a few lines.
• What the node adds — only this is said to the person: the tests run on this computer and their scripts stay in git; the research outlives any window — a letter and a clean window, never a compaction (core step 700); a page per research on the person's own domain, by their yes, and a dashboard behind sign-in; a clock for re-checks after a yes; every research in one data space and one graph (the data, vectors and graph elements), so researches are compared with one yardstick; other hands take a finished dossier over i2i; every hand works in its own window of up to a million tokens while the core keeps only the map.
• The price: a hand is about 150 MB of memory while awake and sleeps between researches; a study of a site through REA and its study browser holds up to about 1 GB while it runs; the strongest model's requests come from the person's subscription.
• A research is not a study. A study takes a product apart to build an analogue, in the element that will build it (study-for-element, its study pass); a research understands, compares and checks, and its dossier is what that element starts from.
• Research that recurs learns (force 6): a question asked again, a re-check, a comparison with the last one all need what was kept.

How the work runs once the shape is chosen — one link at a time, each after the person's yes

1. The question and the goal, in the person's words, verbatim into current-steps.md.
2. Whatever is new is born with its horizon and its starting vector (create-element). The vector names its face — a public page per research, published by the person's yes, and an index of the published ones; its inner pages — a dashboard per research for the architect and the roles the person names: the question, the sources with their dates, the probes and what they showed, the verdict, the open questions, the next re-check; its loads — a static public layer, dashboards over its own tables, sleep between researches; and its door for the neighbours — what is known about a product, answered from its tables and the graph with no model.
3. The record of a research: the question, the sources with dates, the probes and their output, the verdict, what was not checked and why, the date of the next re-check.
4. The probe runs in the hand that holds the research, with a test key the person gives to that hand.
5. The answer to the person — the verdict first, then the evidence, then what was not checked; a page only after their yes.
6. Integration — the element that will integrate takes the dossier over i2i (route-to-element); which hand builds it, the core's map says.
7. Community skills for research — skills_find first; a skill goes into the hand that does the work, after the person's yes.

Laws

• Every probe that spends money or calls an outside service, every publication and every rhythm of re-checks starts after the person's yes.
• No claim about a foreign service without its page and its date; no «nobody else does this»; what is on the way is said as on the way.
• The keys of an outside service live only in the hand that uses them; personal data never goes into git.
• An outside order for a research (a2a) is served by the hand's door, never by a Claude session on the owner's subscription (provider-rules); a2a is a plan until the Lagrix network lives.